Why Agentforce
Give the foundation a consistent way to deliver approved activities, maintain information, and connect families with people who can help
The case for Agentforce is operational before it is educational, and the product names matter far less than the problem they would solve. A foundation running this across schools and events needs one consistent way to deliver approved activities, keep resource information current, and connect a caregiver with a person when the moment calls for one. That is a content, routing, and record-keeping problem, and it is the part that quietly decides whether a promising activity survives contact with a real program calendar. One reviewed activity library, one support queue, and one set of records can serve EAT, LEARN, and PLAY at once, which is why the same setup does not have to be built three times.
Approved content
Retrieve educator-reviewed activities and current resource information, so a family gets the same reviewed material whether they arrive through a school event or a library partner.
Bounded actions
Keep the Coach within specific learning and support tasks: run an activity, look up an approved record, hand off to a person. Nothing else is in scope.
Human assistance
Route a request to designated staff with the context already attached, so a caregiver does not repeat themselves and staff know which activity and which record are involved.
Evaluation
Track minimal participation and feedback information — enough to assess usefulness and access across languages and modes, and no more than that.
The same four capabilities apply across EAT, LEARN, and PLAY, and across whichever channel a family already uses. One reviewed activity library, one set of resource records, and one support queue can serve a book event, a schoolyard program, and a food partner without rebuilding the operation three times.
Agentforce is worth piloting if these operational benefits justify the implementation effort and ongoing costs. The pilot should compare AI-supported activities with equivalent printed activities.
This proposal does not claim that the foundation currently uses Salesforce, and this website is not connected to Agentforce.
Proposed architectureReviewer detail — a proposed design, not a built system.
Where a family starts
iOS and Android app
A family app that calls Agentforce behind the scenes. No separate chat product to learn.
A messaging channel many Oakland families already use every day.
Apple Messages for Business
Activities delivered inside the iPhone Messages app a caregiver already has open.
Agentforce Voice
A spoken conversation over a phone call, for caregivers who would rather talk than type, or who are cooking, driving, or away from a screen. Same consent gate, same output checks before anything is heard.
On-device translation
On the app path only: the phone handles language conversion locally, so fewer paid model calls are needed. An efficiency idea to test, not a measured saving.
Consent and sensitive-input filtering
Every channel passes through the same gate. Sensitive details are filtered before any model sees them.
Salesforce messaging and routing
One place where all channels converge, so a family gets the same activity whichever door they came through.
Agentforce (headless)
The guided activity runs here, with no interface of its own — each channel supplies the presentation.
Approved content library
Educator-reviewed activities and current resource records, retrieved through a Data 360-backed Agentforce Data Library. No open-web retrieval.
Output checks
Replies are reviewed against the approved activity before a family receives them.
Caregiver and child
The activity arrives back in the same conversation the family started in.
Human-support route
At any step the Coach can hand the conversation to designated foundation or partner staff, carrying the activity and record context with it.
Minimal evaluation-data route
A separate, narrow path recording participation and feedback only, kept apart from the activity conversation.
Voice runs the same route as the other channels: nothing is spoken back to a family until it has passed output checks. It also carries its own questions to test — how audio is handled, how long transcripts are kept, and how well speech recognition performs across accents, background noise, and Spanish.
This is a proposed design. Nothing shown here is built, connected, or approved, and each channel still needs technical, accessibility, coverage, and cost validation.
- Salesforce could manage caregiver consent, preferences, and support cases.
- Data 360-backed retrieval could provide approved content.
- Headless Agentforce could guide the bounded activity behind whichever channel the family uses.
- An iOS and Android app, WhatsApp, Apple Messages for Business, and Agentforce Voice are the proposed delivery channels, subject to technical, accessibility, coverage, and cost validation.
- Agentforce Voice is the accessibility-motivated channel — for caregivers with limited reading comfort or low vision, hands-free moments, and shared devices — but speech recognition accuracy across accents and in Spanish must be tested before relying on it.
- Voice audio and transcripts are treated as sensitive input, filtered like any other channel, with no child voice capture.
- On-device translation on the app path may reduce paid model usage, but translation quality across languages must be checked before relying on it.
- Start with one channel.
- No open-web retrieval in the proposed pilot.
- No autonomous eligibility or enrollment decisions.
- Sensitive data must be filtered before model access.
- Do not assume automatic Agentforce masking.
- External model retention commitments do not eliminate application or audit logs.